Hyperparameters Analysis of Machine Learning Techniques for Classification of Marine Targets in SAR Images

Due to the extensive coastal area of Brazil, pattern recognition techniques based on artificial intelligence can search for targets at sea faster for surveillance, rescue, or illicit combat activities. This article presents a hyperparameter analysis of machine learning techniques to classify targets in SAR images. We considered a data set with vertical horizontal polarization SAR images from Campos Basin, Rio de Janeiro, to classify oil platforms and ships. The classification attributes are extracted through a convolutional neural network VGG-16 pre-trained with the ImageNet data set. Then, four machine learning techniques are evaluated, random forest, decision tree, k-nearest-neighbors, and logistic regression. As a metric for assessing the classifiers, accuracy (Acc) and area under the curve (AUC) are used. The grid search technique is used to identify the best combination of parameters of the classifiers with the highest Acc and AUC. Finally, the best result is the logistic regression classifier.